EgoBody: Human Body Shape and Motion of Interacting People from Head-Mounted Devices

EgoBody dataset is a novel large-scale dataset for egocentric 3D human pose, shape and motions under interactions in complex 3D scenes.

[Project page] [Paper] [Dataset] [EgoBody challenge]

News

[December 04, 2023] Text descriptions for motions are provided by Motion-X Dataset.

[October 20, 2022] All modalities of EgoBody are released (including third-person view RGBD, 3D scene, eye gaze/hand/head tracking, 3D human shape and motion annotations for the camera wearer, etc.)!

[July 17, 2022] The EgoBody challenge is released! The first phase of the challenge will end at October 1st. The participants are welcome to submit a 2-4 page abstract to our ECCV workshop.

[July 08, 2022] The EgoSet (egocentric RGB subset of EgoBody) is released! Other modalities (third-person view RGBD, 3D scene, eye gaze, etc.) will come soon.

[June 01, 2022] The EgoBody dataset will be part of the ECCV2022 workshop: Human Body, Hands, and Activities from Egocentric and Multi-view Cameras. Please check out our workshop website for more information.

Dataset License/Download

Please register, sign the dataset license and download the dataset at https://egobody.inf.ethz.ch.

Dataset Information

EgoBody dataset contains 125 sequences, 36 subjects and 15 indoor scenes. Please find more detailed statistics in our paper.

frames	train	val	test	total
MVSet	116630	29140	73961	219731
EgoSet	105388	25416	68307	199111
EgoSet_interactee	90124	23332	62155	175611

MVSet: synchronized frames captured from the Azure Kinects, from multiple third-person views
EgoSet: egocentric RGB frames captured from the HoloLens, calibrated and synchronized with the Kinect frames
EgoSet_interactee: frames where the interactee is visible in the egocentric view

Dataset Documentation

Info/calibration files:

EgoBody
├── data_info_release.csv
├── data_splits.csv
├── kinect_cam_params
│   ├── kinect_master/kinect_sub_1/kinect_sub_2/kinect_sub_3/kinect_sub_4
│   │   ├── Color.json
│   │   ├── IR.json
├── calibrations
│   ├── RECORDING_NAME
│   │   ├── kinect12_to_world/$scene_name$.json
│   │   ├── holo_to_kinect12.json
│   │   ├── kinect_11to12_color.json
│   │   ├── kinect_13to12_color.json
│   │   ├── (kinect_14to12_color.json)
│   │   ├── (kinect_15to12_color.json)

data_info_release.csv: basic information for all sequences
- recording_name: name for each sequence, named as recording_202xxxxx_Sxx_Sxx_xx. 202xxxxx is the capture date, and Sxx_Sxx refers to subjectID(cemera_wearer)_subjectID(interactee)
- body_idx_0: gender of the body with index 0 in each sequence
- body_idx_1: gender of the body with index 1 in each sequence
- body_idx_fpv: body index and gender of the interactee in each sequence
- start_frame/end_frame: the frame ID of starting/ending frame for each sequence
- scene_name: name of the 3D scene for each sequence
data_splits.csv: train/validation/test sequence splits
calibrations: extrinsics between kinects/hololens/3D scene for each sequence
- Note that the kinects are labelled as: 12 (master), 11 (sub_1), 13 (sub_2), 14 (sub_3), 15 (sub_4)
- kinect12_to_world/$scene_name$.json: extrinsics between master kinect RGB camera and the 3D scene mesh
- holo_to_kinect12.json: extrinsics between master kinect RGB camera and the hololens world coordinate system
- kinect_11to12_color.json: extrinsics between RGB camera of the master kinect and sub_1 kinect
- kinect_13to12_color.json: extrinsics between RGB camera of the master kinect and sub_2 kinect
- kinect_14to12_color.json: extrinsics between RGB camera of the master kinect and sub_3 kinect, only exists for data captured in 2022
- kinect_15to12_color.json: extrinsics between RGB camera of the master kinect and sub_4 kinect, only exists for data captured in 2022
kinect_cam_params:
- kinect_master/kinect_sub_1/kinect_sub_2/kinect_sub_3/kinect_sub_4: intrinsics and extrinsics of the color/depth camera for each kinect

Egocentric data streams

EgoBody
├── egocentric_color
│   ├── RECORDING_NAME
│   │   ├── 202x-xx-xx-xxxxxx
│   │   │   ├── PV
│   │   │   ├── 202x-xx-xx-xxxxxx_pv.txt
│   │   │   ├── keypoints.npz
│   │   │   ├── valid_frame.npz
├── egocentric_depth
│   ├── RECORDING_NAME
│   │   ├── 202x-xx-xx-xxxxxx
│   │   │   ├── ...
├── egocentric_gaze
│   ├── RECORDING_NAME
│   │   ├── 202x-xx-xx-xxxxxx
│   │   │   ├── 202x-xx-xx-xxxxxx_head_hand_eye.csv

egocentric_color: egocentric RGB images and hololens camera information
- RECORDING_NAME: recording_name
- PV: egocentric RGB frames of the current sequence, named as timestamp_frame_xxxxx.jpg, where frame_xxxxx is the ID for each frame
- 202x-xx-xx-xxxxxx_pv.txt:
  - row 1: RGB(PV) camera cx, cy, w, h of the current sequence
  - row >=2: timestamp, fx, fy, pv2world_transform of each each RGB(pv) frame. pv2world_transform is the extrinsics between the RGB(PV) camera of each frame and the hololens world coordinate of the current sequece (each hololens sequence has a consitent world coordinate system for the whle sequence).
- keypoints.npz:
  - imgname: egocentric PV image paths (e.x., egocentric_color/RECORDING_NAME/202x-xx-xx-xxxxxx/PV/timestamp_frame_xxxxx.jpg)
  - center: center of the bounding box (to crop the person out) for each PV frame
  - scale: scale of the bounding box for each PV frame
  - keypoints: openpose body joints (BODY_25 format) of the person (interactee) for each PV frame
  - gender: gender of the interactee for each PV frame
- valid_frame.npz:
  - imgname: egocentric PV image paths (e.x., egocentric_color/RECORDING_NAME/202x-xx-xx-xxxxxx/PV/timestamp_frame_xxxxx.jpg)
  - valid: True/False, True indicates that the detected openpose body joints >= 6 for the interactee in each PV frame
- Note: keypoints.npz, valid_frame.npz and 202x-xx-xx-xxxxxx_pv.txt can contain frame IDs outside of the range of [start_frame, end_frame], please ignore those frames.
egocentric_depth: egocentric depth recordings
- RECORDING_NAME: recording_name
- each sequence contains recorded depth, lookup table, and depth camera extrinsics (please refer here for more information)
egocentric_gaze: egocentric eye gaze recordings
- RECORDING_NAME: recording_name
- 202x-xx-xx-xxxxxx_head_hand_eye.csv: each row includes timestamp and head/hand/eye gaze tracking for the current timestamp, please refer load_head_hand_eye_data() in utils.py for details.

Third-person view data streams

EgoBody
├── kinect_color
│   ├── RECORDING_NAME
│   │   ├── master/sub_1/sub_2(/sub_3/sub_4)
│   │   │   ├── frame_xxxxx.jpg
├── kinect_depth
│   ├── RECORDING_NAME
│   │   ├── master/sub_1/sub_2(/sub_3/sub_4)
│   │   │   ├── frame_xxxxx.png

kinect_color: multi-view third-person view RGB images captured by Kinect cameras
- RECORDING_NAME: recording_name
- master/frame_xxxxx.jpg: RGB frame for master kinect
- sub_1/frame_xxxxx.jpg: RGB frame for sub_1 kinect ...
- Note that here frame ID frame_xxxxx is synchronized with the corresponding egocentric RGB frame timestamp_frame_xxxxx.jpg of the same sequence.
kinect_depth: multi-view third-person view depth images captured by Kinect cameras
- RECORDING_NAME: recording_name
- master/frame_xxxxx.png: depth frame for master kinect
- sub_1/frame_xxxxx.png: depth frame for sub_1 kinect ...
- Note that here frame ID frame_xxxxx is synchronized with the kinect RGB frame with frame ID frame_xxxxx

3D scene meshes

EgoBody
├── scene_mesh
│   ├── $scene_name$
│   │   ├── $scene_name$.obj

$scene_name.obj$ : 3D scene mesh for scene $scene_name$

3D human pose, shape and motion annotations

EgoBody
├── smplx_interactee_train
│   ├── RECORDING_NAME/body_idx_x/results/frame_xxxxx/000.pkl
├── smplx_interactee_val
├── smplx_camera_wearer_train
│   ├── RECORDING_NAME/body_idx_x/results/frame_xxxxx/000.pkl
├── smplx_camera_wearer_val
├── smpl_interactee_train
│   ├── RECORDING_NAME/body_idx_x/results/frame_xxxxx/000.pkl
├── smpl_interactee_val
├── smpl_camera_wearer_train
│   ├── RECORDING_NAME/body_idx_x/results/frame_xxxxx/000.pkl
├── smpl_camera_wearer_val

smplx_interactee_train/val and smplx_camera_wearer_train/val: SMPL-X body parameters for each frame of the interactee/camera wearer in training/val set
- always in the coordinate system of the master kinect RGB camera
- body_idx_x is the body index of the interactee/camera wearer in the current sequence, and frame_xxxxx is the ID for each frame.
smpl_interactee_train/val and smpl_camera_wearer_train/val: SMPL body parameters for each frame of the interactee/camera wearer

Motion text descriptions

Text descriptions for motions are provided by Motion-X Dataset.

Visualization Code

Render SMPL-X/SMPL bodies from the egocentric (hololens) view:

python release_renderer_fpv_gaze.py --release_data_root=PATH/TO/DATASET  --save_root=PATH/TO/SAVE/RESULTS --recording_name RECORDING_NAME --scene_name SCENE_NAME

Available options:

model_type: smpl/smplx, render SMPL-X or SMPL bodies
plot_2d_joints: if set to True, plot 2D joints of openpose detections and the projected 2D joints of ground truth SMPL-X/SMPL bodies
plot_gaze: if set to True, plot 2D projection of the camera wearer's gaze point on egocentric view images
rendring_mode: body renders 3D body mesh projected on the RGB images, 3d renders 3D body mesh in 3D scenes, both renders both options
model_folder: the path to SMPL-X/SMPL models

Render SMPL-X/SMPL bodies from the third-person (kinect) view:

python release_renderer_kinect.py --release_data_root=PATH/TO/DATASET  --save_root=PATH/TO/SAVE/RESULTS --recording_name RECORDING_NAME --scene_name SCENE_NAME

Available options:

model_type: smpl/smplx, render SMPL-X or SMPL bodies
view: from which view of kinect to render the body (options: master/sub_1/sub_2/sub_3/sub_4)
rendring_mode: body renders 3D body mesh projected on the RGB images, 3d renders 3D body mesh in 3D scenes, both renders both options
model_folder: the path to SMPL-X/SMPL models

Visualize point clouds from kinect RGB/depth and the 3d scene mesh together:

python release_vis_kinect_scene.py --release_data_root=PATH/TO/DATASET  --recording_name RECORDING_NAME --scene_name SCENE_NAME

Available options:

vis_frame_id: which frame to visualize, in the format of xxxxx, for example, `03000'

Visualize point clouds from kinect RGB/depth of all kinect views together:

python release_vis_kinect_pcd.py --release_data_root=PATH/TO/DATASET  --recording_name RECORDING_NAME

Available options:

vis_frame_id: which frame to visualize, in the format of xxxxx, for example, `03000'

To read the depth, head/hand tracking data:

Please refer to HoloLens2ForCV for details.

To read hololens depth and convert to point clouds, use: HoloLens2ForCV/Samples/StreamRecorder/StreamRecorderConverter/save_pclouds.py
To read head/hand tracking data and project onto the egocentric image, use: HoloLens2ForCV/Samples/StreamRecorder/StreamRecorderConverter/project_hand_eye_to_pv.py

Baseline Results

Method	MPJPE	PA-MPJPE	V2V	PA-V2V
CMR	200.7	109.6	218.7	136.8
SPIN	182.8	116.6	187.3	123.8
LGD	158.0	99.9	168.3	106.0
METRO	153.1	98.4	164.6	106.5
PARE	123.0	83.8	131.4	89.7
EFT	123.9	78.4	135.0	86.0
SPIN-ft	106.5	67.1	120.9	78.3
METRO-ft	98.5	66.9	110.5	76.8
EFT-ft	102.1	64.8	116.1	74.8

Here '-ft' denotes results of fine-tuning SPIN, METRO and EFT on our training set.

Citation

@inproceedings{Zhang:ECCV:2022,
   title = {EgoBody: Human Body Shape and Motion of Interacting People from Head-Mounted Devices},
   author = {Zhang, Siwei and Ma, Qianli and Zhang, Yan and Qian, Zhiyin and Kwon, Taein and Pollefeys, Marc and Bogo, Federica and Tang, Siyu},
   booktitle = {European conference on computer vision (ECCV)},
   month = oct,
   year = {2022}
}

Acknowledgments

This work was supported by the Microsoft Mixed Reality & AI Zurich Lab PhD scholarship. Qianli Ma is partially funded by the Max Planck ETH Center for Learning Systems. We sincerely thank Francis Engelmann, Korrawe Karunratanakul, Theodora Kontogianni, Qi Ma, Marko Mihajlovic, Sergey Prokudin, Matias Turkulainen, Rui Wang , Shaofei Wang and Samokhvalov Vyacheslav for helping with the data capture and processing, Xucong Zhang for the discussion of data collection and Jonas Hein for the discussion of the hardware setup. Siyu Tang acknowledges the SNF grant 200021 204840.

Relevant projects

The motion reconstruction pipeline benefits from:

Learning Motion Priors for 4D Human Body Capture in 3D Scenes (ICCV 2021 (Oral))
Siwei Zhang, Yan Zhang, Federica Bogo, Marc Pollefeys and Siyu Tang

Resolving 3D Human Pose Ambiguities with 3D Scene Constraints (ICCV 2019)
Mohamed Hassan, Vassilis Choutas, Dimitrios Tzionas and Michael J. Black

sanweiliti/EgoBody

EgoBody: Human Body Shape and Motion of Interacting People from Head-Mounted Devices

News

Dataset License/Download

Dataset Information

Dataset Documentation

Info/calibration files:

Egocentric data streams

Third-person view data streams

3D scene meshes

3D human pose, shape and motion annotations

Motion text descriptions

Visualization Code

Baseline Results

Citation

Acknowledgments

Relevant projects